Machine Learning for Large-Scale Quality Control of 3D Shape Models in\n Neuroimaging

As very large studies of complex neuroimaging phenotypes become more common,\nhuman quality assessment of MRI-derived data remains one of the last major\nbottlenecks. Few attempts have so far been made to address this issue with\nmachine learning. In this work, we optimize predictive models of quality for\nmeshes representing deep brain structure shapes. We use standard vertex-wise\nand global shape features computed homologously across 19 cohorts and over 7500\nhuman-rated subjects, training kernelized Support Vector Machine and Gradient\nBoosted Decision Trees classifiers to detect meshes of failing quality. Our\nmodels generalize across datasets and diseases, reducing human workload by\n30-70\\%, or equivalently hundreds of human rater hours for datasets of\ncomparable size, with recall rates approaching inter-rater reliability.\n